Ouzhang's picture
Add files using upload-large-folder tool
31dc8dc verified
Raw
History Blame Contribute Delete
21.1 kB
import argparse, ast, io, json, os, sys, time, textwrap, multiprocessing as mp
import concurrent.futures as cf
from pathlib import Path
import math
import numpy as np
from termcolor import cprint
from tqdm import tqdm
from omegaconf import DictConfig, ListConfig, OmegaConf, MISSING
def get_config():
cli_conf = OmegaConf.from_cli()
yaml_conf = OmegaConf.load(cli_conf.config)
return OmegaConf.merge(yaml_conf, cli_conf)
from concurrent.futures import as_completed
import textwrap
def _run_many_pipe(snippet: str, tests: list[str], conn):
import textwrap
import sys
import traceback
results = []
debug_info = []
try:
ns = {}
# Normalize line endings to \n before dedent and execution
normalized_snippet = snippet.replace('\r\n', '\n').replace('\r', '\n')
dedented_snippet = textwrap.dedent(normalized_snippet)
debug_info.append(f"Original snippet (repr): {repr(snippet[:200])}")
debug_info.append(f"Normalized snippet (repr): {repr(normalized_snippet[:200])}")
debug_info.append(f"Dedented snippet (repr): {repr(dedented_snippet[:200])}")
# Execute the snippet
exec(dedented_snippet, ns, ns)
debug_info.append(f"Functions in namespace after exec: {[k for k in ns.keys() if not k.startswith('__')]}")
# Run tests
for idx, stmt in enumerate(tests):
try:
exec(stmt, ns, ns)
results.append(True)
debug_info.append(f"Test {idx+1}/{len(tests)} '{stmt}' PASSED")
except SystemExit:
results.append(True)
debug_info.append(f"Test {idx+1}/{len(tests)} '{stmt}' PASSED (SystemExit)")
except AssertionError as e:
results.append(False)
debug_info.append(f"Test {idx+1}/{len(tests)} '{stmt}' FAILED (AssertionError): {e}")
except Exception as e:
results.append(False)
debug_info.append(f"Test {idx+1}/{len(tests)} '{stmt}' FAILED ({type(e).__name__}): {e}")
debug_info.append(f" Traceback: {traceback.format_exc()}")
# Send results with debug info
conn.send((results, debug_info))
except SystemExit as e:
debug_info.append(f"Snippet execution SystemExit: {e}")
conn.send(([True] * len(tests), debug_info))
except Exception as e:
debug_info.append(f"Snippet execution FAILED ({type(e).__name__}): {e}")
debug_info.append(f" Traceback: {traceback.format_exc()}")
conn.send(([False] * len(tests), debug_info))
finally:
try: conn.close()
except Exception: pass
def _check_snippet_many(snippet: str, tests: list[str], t_limit: int,
spawn_slack: float = 5.0, debug_file=None) -> tuple[list[bool], list[str]]:
import time, multiprocessing as mp
import os, sys
ctx = mp.get_context("spawn")
parent_conn, child_conn = ctx.Pipe(duplex=False)
p = ctx.Process(target=_run_many_pipe, args=(snippet, tests, child_conn), daemon=True)
# Start process and close child connection
try:
p.start()
except Exception as e:
child_conn.close()
parent_conn.close()
return ([False] * len(tests), [f"Failed to start process: {e}"])
child_conn.close()
# Increase timeout: t_limit for execution + spawn_slack for process startup/overhead
deadline = time.monotonic() + t_limit + spawn_slack
res = None
debug_info = []
process_exited = False
try:
# Poll for result with timeout
while time.monotonic() < deadline:
if parent_conn.poll(0.1): # Check every 100ms
try:
res = parent_conn.recv()
break
except (EOFError, OSError) as e:
# Connection closed or error
debug_info.append(f"Connection error: {e}")
res = None
break
# Check if process has exited
if not p.is_alive():
process_exited = True
# Try to get result one more time
if parent_conn.poll(0.1):
try:
res = parent_conn.recv()
break
except (EOFError, OSError):
pass
break
# Final attempt to get result
if res is None and parent_conn.poll(0.1):
try:
res = parent_conn.recv()
except (EOFError, OSError):
pass
# If no result, mark as timeout
if res is None:
if p.is_alive():
try:
p.terminate()
p.join(timeout=1.0)
if p.is_alive():
p.kill()
except Exception:
pass
timeout_msg = "Timeout or no response from child process"
if process_exited:
timeout_msg += " (process exited without sending result)"
res = ([False] * len(tests), [timeout_msg])
else:
# Handle both old format (list) and new format (tuple)
if isinstance(res, tuple):
res, debug_info = res
else:
debug_info = ["Legacy format: no debug info"]
except Exception as e:
# Unexpected error
try:
if p.is_alive():
p.terminate()
p.join(timeout=1.0)
except Exception:
pass
res = ([False] * len(tests), [f"Unexpected error in _check_snippet_many: {e}"])
debug_info = []
finally:
# Cleanup
try:
if p.is_alive():
p.terminate()
p.join(timeout=1.0)
except Exception:
pass
try:
parent_conn.close()
except Exception:
pass
# Write debug info to file if provided
if debug_file:
try:
# Ensure directory exists
debug_dir = os.path.dirname(debug_file)
if debug_dir:
os.makedirs(debug_dir, exist_ok=True)
with open(debug_file, 'a', encoding='utf-8') as f:
f.write("=" * 80 + "\n")
f.write(f"Snippet (first 200 chars): {repr(snippet[:200])}\n")
f.write(f"Tests: {tests}\n")
f.write(f"Results: {res}\n")
for line in debug_info:
f.write(line + "\n")
f.write("=" * 80 + "\n\n")
except Exception as e:
# Don't fail silently - at least print to stderr
import sys
print(f"Warning: Could not write to debug file {debug_file}: {e}", file=sys.stderr)
return ([bool(x) for x in res], debug_info)
from concurrent.futures import ThreadPoolExecutor, as_completed
def evaluate_function_dataset(data: list[dict], n_workers: int | None = None, debug_file: str | None = None):
import os
import time
n_cpu = os.cpu_count() or 4
n_workers = max(1, int(n_workers)) if n_workers is not None else n_cpu
# Create debug file if not provided
if debug_file is None:
debug_file = "./tmp/execute_debug.log"
# Ensure directory exists
debug_dir = os.path.dirname(debug_file)
if debug_dir:
os.makedirs(debug_dir, exist_ok=True)
# Clear debug file at start
try:
with open(debug_file, 'w', encoding='utf-8') as f:
f.write(f"Debug log started at {time.strftime('%Y-%m-%d %H:%M:%S')}\n\n")
print(f"Debug log file: {os.path.abspath(debug_file)}")
except Exception as e:
print(f"Warning: Could not create debug file {debug_file}: {e}", file=sys.stderr)
for item in data:
m_code = len(item["extracted_output"])
m_test = len(item["test_list"])
item["execution_result"] = [[None] * m_test for _ in range(m_code)]
item["correctness"] = [[False] * m_test for _ in range(m_code)]
item.setdefault("step_map", [])
tasks = []
for idx, item in enumerate(data):
t_limit = item.get("test_time_limit", 1)
tests = item["test_list"]
for i, snippet in enumerate(item["extracted_output"]):
tasks.append((idx, i, snippet, tests, t_limit))
futures = {}
from tqdm.auto import tqdm
with ThreadPoolExecutor(max_workers=n_workers) as pool, \
tqdm(total=len(tasks)*len(data[0]["test_list"]), desc=f"Function tests ({n_workers} threads)",
dynamic_ncols=True, mininterval=0.1, miniters=1) as pbar:
for idx, i, snippet, tests, t_limit in tasks:
fut = pool.submit(_check_snippet_many, snippet, tests, t_limit, debug_file=debug_file)
futures[fut] = (idx, i)
for fut in as_completed(futures):
idx, i = futures[fut]
try:
result = fut.result()
# Handle both old format (list) and new format (tuple)
if isinstance(result, tuple):
ok_list, debug_info = result
# Write debug info to file for failed cases
if any(not ok for ok in ok_list):
try:
with open(debug_file, 'a', encoding='utf-8') as f:
f.write(f"\n[FAILED CASE] Item {idx}, Code {i}\n")
f.write(f"Tests: {data[idx]['test_list']}\n")
f.write(f"Results: {ok_list}\n")
for line in debug_info:
f.write(line + "\n")
except Exception:
pass
else:
ok_list = result
except Exception as e:
ok_list = [False] * len(data[idx]["test_list"])
try:
with open(debug_file, 'a', encoding='utf-8') as f:
f.write(f"\n[EXCEPTION] Item {idx}, Code {i}: {e}\n")
except Exception:
pass
# Ensure ok_list is a list and matches the expected number of tests
expected_tests = len(data[idx]["test_list"])
if not isinstance(ok_list, list):
ok_list = [False] * expected_tests
try:
with open(debug_file, 'a', encoding='utf-8') as f:
f.write(f"\n[TYPE ERROR] Item {idx}, Code {i}: "
f"ok_list is not a list, got {type(ok_list)}\n")
except Exception:
pass
elif len(ok_list) != expected_tests:
# Log the mismatch for debugging before fixing
original_len = len(ok_list)
try:
with open(debug_file, 'a', encoding='utf-8') as f:
f.write(f"\n[LENGTH MISMATCH] Item {idx}, Code {i}: "
f"Expected {expected_tests} tests, got {original_len} results\n")
except Exception:
pass
# Pad or truncate to match expected length
if len(ok_list) < expected_tests:
ok_list = ok_list + [False] * (expected_tests - len(ok_list))
else:
ok_list = ok_list[:expected_tests]
# Ensure execution_result and correctness lists are properly initialized
if i >= len(data[idx]["execution_result"]):
# Extend if needed
while len(data[idx]["execution_result"]) <= i:
data[idx]["execution_result"].append([None] * expected_tests)
data[idx]["correctness"].append([False] * expected_tests)
elif len(data[idx]["execution_result"][i]) != expected_tests:
# Resize if needed
data[idx]["execution_result"][i] = [None] * expected_tests
data[idx]["correctness"][i] = [False] * expected_tests
for j, ok in enumerate(ok_list):
if j < len(data[idx]["execution_result"][i]):
data[idx]["execution_result"][i][j] = bool(ok)
data[idx]["correctness"][i][j] = bool(ok)
pbar.update(1)
return data
def worker_stdio(script, input_val, output_queue):
# Create an iterator over the input lines.
input_lines = iter(input_val.splitlines())
# Override the input() function in the exec context.
def fake_input(prompt=""):
try:
return next(input_lines)
except StopIteration:
raise EOFError("No more input")
# Redirect sys.stdout to capture printed output.
stdout_capture = io.StringIO()
original_stdout = sys.stdout
original_stdin = sys.stdin # Save original stdin
sys.stdout = stdout_capture
sys.stdin = io.StringIO(input_val) # Simulate stdin with input_val
context = {
"__name__": "__main__", # Ensures that `if __name__ == "__main__": ...` will fire
"input": fake_input
}
try:
exec(script, context)
printed_output = stdout_capture.getvalue()
output_queue.put(printed_output)
except SystemExit:
printed_output = stdout_capture.getvalue()
output_queue.put(printed_output)
except Exception as e:
output_queue.put(f"error: {e}")
finally:
sys.stdout = original_stdout
sys.stdin = original_stdin
def run_scripts_with_timeout(scripts, inputs, time_limits, worker, batch_size=100):
results = [None] * len(scripts)
n = len(scripts)
# Process in batches to avoid exhausting file descriptors
for batch_start in range(0, n, batch_size):
batch_end = min(batch_start + batch_size, n)
processes = []
queues = []
deadlines = []
for i in range(batch_start, batch_end):
q = mp.Queue()
p = mp.Process(target=worker, args=(scripts[i], inputs[i], q))
processes.append(p)
queues.append(q)
p.start()
deadlines.append(time.time() + time_limits[i])
while any(p.is_alive() for p in processes):
now = time.time()
for idx, p in enumerate(processes):
i = batch_start + idx
if p.is_alive() and now >= deadlines[idx]:
p.terminate()
results[i] = "Timeout Error"
time.sleep(0.001)
for idx, p in enumerate(processes):
i = batch_start + idx
if results[i] is None:
try:
results[i] = queues[idx].get_nowait()
except Exception as e:
results[i] = f"Execution Error: {e}"
# Clean up processes
for p in processes:
try:
p.join(timeout=0.1)
except Exception:
pass
return results
def test_if_eq(x, y):
return " ".join(x.split()) == " ".join(y.split())
def get_chunk_indices(n, num_chunks):
size, rem = divmod(n, num_chunks)
idx, start = [], 0
for i in range(num_chunks):
extra = 1 if i < rem else 0
end = start + size + extra
idx.append((start, end)); start = end
return idx
from tqdm import tqdm
def run_scripts_with_chunk(code_list, test_input_list, time_limit_list,
worker, num_chunks):
chunks = get_chunk_indices(len(code_list), num_chunks)
exe_results = []
pbar = tqdm(total=len(code_list), desc=f"STDIO tests ({num_chunks} ch)")
for start, end in chunks:
sub_code_list = code_list[start:end]
sub_test_input_list = test_input_list[start:end]
sub_time_limit_list = time_limit_list[start:end]
sub_exe_results = run_scripts_with_timeout(
sub_code_list,
sub_test_input_list,
sub_time_limit_list,
worker
)
exe_results.extend(sub_exe_results)
pbar.update(end - start)
pbar.close()
return exe_results
def evaluate_stdio_dataset(data: list[dict], num_chunks: int):
idx_code, idx_case = [], []
code_list, inp_list, tl_list = [], [], []
for idx, item in enumerate(data):
tl = item.get("test_time_limit", 1)
m_code = len(item["extracted_output"])
m_case = len(item["test_input"])
data[idx]["execution_result"] = [[] for _ in range(m_code)]
data[idx]["correctness"] = [[] for _ in range(m_code)]
item.setdefault("step_map", [])
for c_idx, code in enumerate(item["extracted_output"]):
for k in range(m_case):
idx_code.append((idx, c_idx))
idx_case.append(k)
code_list.append(code)
inp_list.append(item["test_input"][k])
tl_list.append(tl)
exe_results = run_scripts_with_chunk(
code_list, inp_list, tl_list, worker_stdio, num_chunks
)
for i, res in enumerate(exe_results):
idx, c_idx = idx_code[i]
k = idx_case[i]
item = data[idx]
while len(item["execution_result"][c_idx]) < k + 1:
item["execution_result"][c_idx].append("")
item["correctness"][c_idx].append(False)
item["execution_result"][c_idx][k] = res
exp_out = item["test_output"][k]
item["correctness"][c_idx][k] = test_if_eq(res, exp_out)
return data
def main():
cfg = get_config()
project_name = cfg.experiment.project
# Get checkpoint path: evaluation.checkpoint_path > model.pretrained_model > config.model (string)
checkpoint_path = OmegaConf.select(cfg, "evaluation.checkpoint_path", default=MISSING)
if checkpoint_path is MISSING or checkpoint_path is None:
# Support both config.model (string) and config.model.pretrained_model (dict)
if isinstance(cfg.model, str):
pretrained_model = cfg.model
else:
pretrained_model = cfg.model.pretrained_model
else:
pretrained_model = checkpoint_path
# Validate that pretrained_model is set
if pretrained_model is None:
raise ValueError(
"pretrained_model is None. Please set either:\n"
" - evaluation.checkpoint_path in config, or\n"
" - model (as string) or model.pretrained_model in config"
)
dataset = OmegaConf.select(cfg, "evaluation.eval_dataset", default=None)
if dataset is None:
raise ValueError("evaluation.eval_dataset is required in config")
outputs_name = "eval-" + pretrained_model.split("/")[-1] + "-" + dataset
output_base = OmegaConf.select(cfg, "experiment.output_dir", default=None) or getattr(cfg.experiment, "output_dir", "..")
output_base = os.path.expanduser(str(output_base))
_script_dir = os.path.dirname(os.path.abspath(__file__))
if not os.path.isabs(output_base):
output_base = os.path.normpath(os.path.join(_script_dir, output_base))
num_node = cfg.evaluation.num_node
node_index = cfg.evaluation.node_index
if num_node > 1:
file_name = os.path.join(output_base, project_name, "temp_data", f"outputs-{node_index}-{outputs_name}.json")
else:
file_name = os.path.join(output_base, project_name, "temp_data", f"outputs-{outputs_name}.json")
with open(file_name, "r", encoding="utf-8") as f:
data = json.load(f)
func_items = [itm for itm in data if itm.get("test_method","function") == "function"]
stdio_items = [itm for itm in data if itm.get("test_method") == "stdio"]
# --- 1) function ---
if func_items:
updated_func = evaluate_function_dataset(func_items, n_workers=cfg.execute.num_chunk)
func_iter = iter(updated_func)
for i,it in enumerate(data):
if it.get("test_method","function") == "function":
data[i] = next(func_iter)
# --- 2) stdio ---
if stdio_items:
total_scripts = sum(len(it["extracted_output"]) for it in stdio_items)
num_chunks = max(1, math.ceil(total_scripts / cfg.execute.num_chunk))
updated_stdio = evaluate_stdio_dataset(stdio_items, num_chunks=num_chunks)
it_stdio = iter(updated_stdio)
for i, it in enumerate(data):
if it.get("test_method") == "stdio":
data[i] = next(it_stdio)
# --- save JSON ---
os.makedirs(os.path.dirname(file_name), exist_ok=True)
with open(file_name, "w", encoding="utf-8", errors="surrogatepass") as f:
json.dump(data, f, indent=2, ensure_ascii=False)
if __name__ == "__main__":
mp.set_start_method("spawn", force=True)
main()